The question circulating in data science circles right now: are AI agents coming for the job, or coming for the boring parts of it? After watching practitioners work with these tools for the past year, the honest answer is both - and the split is happening faster than most teams expected.
The Tasks That Are Already Gone
Exploratory data analysis - loading a dataset, checking distributions, spotting outliers, generating summary statistics - is largely automatable today. A data scientist who spent 40% of their week on this work in 2024 is spending maybe 15% on it now, with tools like Claude or ChatGPT writing the pandas code on demand. Feature engineering for structured tabular data follows a similar pattern. Writing SQL queries to pull training data? Essentially solved.
This is not a future concern. It's happening in active data teams at companies of all sizes. The repetitive, code-heavy, pattern-following work is compressing.
What Agents Still Can't Do
Here's where it gets specific. AI agents are good at tasks with a clear right answer verifiable against data. They're poor at the upstream work: deciding which question is worth asking in the first place, pushing back on a stakeholder's framing when the metric they want to optimize will produce perverse outcomes, or knowing when a statistically significant result is operationally meaningless.
A concrete example: an agent can train 50 models and report which has the lowest validation loss. It cannot tell you that the metric your product team handed you as a target is a proxy for engagement that actually decreases user retention over 90 days. That requires context about the business, relationships with the people who built the product, and a willingness to have an uncomfortable conversation.
Model deployment and monitoring add another layer. AI agents can flag when a model's predictions drift from baseline - a process called model drift detection, where the model's outputs gradually stop matching real-world outcomes as conditions change. Diagnosing why the drift happened, tracing it to a data pipeline change or a real-world behavioral shift, still requires a human who understands the full stack.
What This Means for Hiring and Skill Prioritization
The data science roles being eliminated are junior positions built around execution: write the query, clean the data, build the dashboard. The roles being created - or more accurately, the roles becoming more valuable - are the ones centered on problem framing and stakeholder translation.
This creates a real squeeze. Junior data scientists traditionally learned judgment by doing the execution work first. If that work disappears, the pipeline for developing senior-level thinking gets disrupted. Teams that are cutting junior headcount to save on costs while relying on AI agents may be quietly undermining their own talent development.
For working data scientists, the practical move is straightforward: stop competing with agents on tasks they win at, and get better at the parts they don't do well. That means communication skills, business context, and the ability to spot a bad problem statement before spending three weeks solving it. The tools are shifting what leverage looks like - but they haven't changed what good judgment is worth.